A Leg Configuration Sensory System for Dynamical Body State

Estimates in a Hexapod Robot∗

Pei-Chun Lin§ Haldun Komsuoḡlu † Daniel E. Kodistchek†
pclin@umich.edu hkomsuog@umich.edu kod@eecs.umich.edu
§Department of Mechanical Engineering, The University of Michigan, Ann Arbor
†Department of Electrical Engineering and Computer Science, The University of Michigan, Ann Arbor

Abstract the availability of full body state estimates [5] through-

We report on a novel leg strain sensory system for
the autonomous robot RHex [8] implemented upon a
cheap, high performance local wireless network [1]. We
introduce a model for RHex’s 4-bar legs [4] relating leg
strain to leg kinematic configuration in the body coor-
dinate frame. We compare against ground truth mea-
surement the performance of the model operating on
real-time leg strain data generated under completely
realistic operating conditions. We introduce an algo-
rithm for computing six degree of freedom body pos-
ture measurements in world frame coordinates from
the outputs of the six leg configuration models, to-
gether with a priori information about the ground. We
discuss the manner in which such stance phase con-
figuration estimates will be fused with other sensory
data to develop the continuous time full body state
estimates for RHex.
out the stance and flight phases of locomotion. This
paper introduces a novel leg strain sensory system for
the autonomous robot RHex [8] that will, in concert
with other sensory modalities, supply the information
stream at the bandwidth required to implement these
new controllers on that machine.
We introduce a model for RHex’s 4-bar legs [4] re-
lating leg strain to leg kinematic configuration in the
body coordinate frame. We compare against ground
truth measurement the performance of the model op-
erating on real-time leg strain data generated under
completely realistic operating conditions. We intro-
duce an algorithm for computing six degree of freedom
body posture in world frame coordinates from the out-
puts of the six leg configuration models, together with
a priori information about the ground. We discuss
the manner in which such stance phase configuration
estimates would be fused with other sensory data to
develop the continuous time full body state estimates.

1 Introduction
The hexapedal robot, RHex [8], presently driven by
task level open loop controllers performs extraordi-
narily well over badly broken and unstable surfaces
at speeds exceeding two body lengths per second [9].
However, both the efficiency and behavioral repertoire
of RHex are limited because of the lack of information
about its state and that of its environment. In our ini-
tial studies with sensor based controllers we observed
significant behavioral improvement from even minimal
feedback [2,6]. In this paper we describe a much longer
step toward the implementation of robust full state es- Figure 1: RHex with 4-bar legs in a natural setting.
timation for RHex.
A recently developed high performance family of dy-
Designed for simplicity, RHex has only one actu-
namical hexapedal robot running controllers presumes
ated degree of freedom per leg. In consequence, the
∗ This work is supported by DARPA/SPAWAR Contract legs cannot generally protract without stubbing the
N66001-00-C-8026. ground, and continuous locomotion requires their con-

1
tinual recirculation. The strain measurements from
the compliant members of the freely rotating 4-bar
legs are obtained through a wireless high speed net-
work, LegNet [1], which is also described briefly.
The organization of the paper is as follows. We start
with a brief introduction to the 4-bar legs in Section 2
followed by a modeling discussion in Section 3 that in-
troduces a polynomial representation of the empirical
function relating strain measurement to leg configura-
tion. Section 4 concerns the data collection and fitting
procedures for computation of leg model. Model per-
formance is demonstrated by emprical data. A body
posture computation algorithm is explained in Section
5. Emprical data is presented to discuss its perfor-
mance. Section 6 concludes with final remarks, a sum-
mary of contributions and future directions.
2 4-Bar Legs
The RHex 4-bar leg [4] illustrated in Figure 2 is a
passive mechanical system composed of four parts: a
hip clamp; front and rear compliant members; and a
shin. The compliant members are thin rectangular
fiber glass strips that are rigidly fixed at one end to
the shin and connected to the hip clamp at the oppo-
site end by a revolute joint (hinge) creating a 4-bar
linkage structure on a plane which will be refer as the
“leg plane” and denoted by L := R2 . The compliant
parts are flexible in directions within the leg plane, L,
but “infinitely” stiff in other directions for all practi-
cal purposes. Combined with the constraints imposed
by the closed chain mechanism the entire leg struc-
ture remains in the leg plane, L, for all times. The
hip clamp rigidly attaches this leg structure to the hip
motor shaft such that the shaft axis is normal to the
leg plane, L, and goes through its origin.
Leg Plane, L
L effectively as a 1-DOF mechanism. As a result, we
Motor Shaft Axis
Hip Clamp
Strain Gauge
Rear Compliant Part
Front Compliant Part
Shin
Toe
Figure 2: Key components of 4-bar leg mechanism.
A strain gauge (Vishay EA-250BK-10C) is installed
2
onto the rear compliant part for measurement of
strain, σ ∈ Σ := R+ , which varies as a function of
leg configuration. The gauge is located on the inside
face near the hinge where strain is the smallest. In our
emprical studies the maximum strain measured during
tripod walking gait is approximately 1000 microstrain
which falls within the 106 cycle at 1500 microstrain
range reported in the strain gauge specifications1 .
RHex’s freely rotating legs introduce a significant
technical difficulty in taking measurements from sen-
sors installed on them — this is essentially a “remote
sensing” problem. In our implementation strain gauge
measurements are transfered to the PC104 stack over a
bi-directional digital wireless communication network,
LegNet [1], featuring 50 Kbaud communication chan-
nel over a 916MHz carrier connecting a master located
in the PC104 stack in the body to six self-contained
embedded slave units mounted on each leg. In this
setting we achieve synchronous sampling of all legs at
300Hz at 6-bit resolution.
3 Leg Model
We prefer to represent leg configuration in the clamp
frame, C := R+ × S 1 — a polar coordinate system,
attached to the hip clamp as depicted in Figure 3(b),
whose states are the distance between hip attachment
point and the toe, l ∈ R+ , or “the effective leg length”;
and, the angular deflection of toe vector from its rest
position, δ ∈ S 1 .
Considering a quasi-static setting we ignore damp-
ing and leg mass and model the 4-bar leg as a pure
passive spring attached to the hip clamp with torsional
and radial compliance giving rise to a memoryless leg
configuration model. Furthermore, since the locus of
the physically valid toe positions, presented in Figure
4(left), has very small area we will consider 4-bar leg
define a scalar leg model mapping the strain, σ, to
the polar coordinates of the toe, c, given by a Taylor
function, m : Σ → C,
 PN £ ¤ 
· ¸ n1j .σ j
l  j=0 
c= = m(σ) :=   (1)
δ  PN £ ¤ 
n2j .σ j
j=0
1 At a typical walking speed, the legs recirulate at roughly
2 Hz, hence we should expect mechanical sensor failures at a
specified leg every 5 x 105 sec, hence, with six operating simul-
taneously, at some leg every 8 x 104 sec. Thus, at a typical a
cruising speed of 1 m/s, we would expect to travel 80 km before
suffering a leg sensor failure.
 B Model Fit Cross Validation
Order p(l, ˆl) p(δ, δ̂) p(l, ˆl) p(δ, δ̂)
C 1 1.12% 4.87% 1.20% 5.44%
l = l0 l 2 1.11% 4.79% 1.16% 5.51%
3 1.05% 4.53% 1.11% 5.19%
4 1.05% 4.43% 1.11% 5.26%

Table 1: Fitting and cross validation percentage nor-

φ malized root mean square error for leg model in (1)
δ
θ=φ at different polynomial orders. Small values indicate
θ
(a) (b) successful model prediction.

Figure 3: 4-bar leg (a) at rest and (b) during com-
pression. Left sketch depicts two coordinate frames:
the body frame, B; and a clamp frame, C. The right
sketch illustrates the basic leg configuration variables,
length, l, and deflection, δ, on a compressed leg. The
dashed line depicts the approximate one dimensional
locus of physically relevant toe positions.
© 1 ª model prediction of a physical measurment, d, we mea-
where the length coefficients,
© 2 n i |i =
ª 0, ..., N , and
deflection coefficients, ni |i = 0, ..., N , are computed
by the fitting studies in Section 4.1.
the set of physically relevant toe positions which forms
a “thin” set — i.e., essentially a thinkened curve — in
the leg plane, L.
Figure 4(center) and Figure 4(right) illustrate
scatter-plots of the effective leg length, l, and the de-
flection angle, δ, versus measured strain, σ, respec-
tively. We compute the coefficients of the leg model
in (1) by ordinary least squares. Letting dˆ denote
sure model performance by the normalized root mean
ˆ between the original data, d, and
square error, p(d, d),
the corresponding model output, d, ˆ given by

4 System Identification
4.1 Data Collection and Model Fitting
Study
A benchtop setup [3] is utilized to collect quasi-static
leg data for modeling. This two degree of freedom
setup allows the user to manually alter the configura-
tion of an attached leg while recording the strain, σ,
effective leg length, l, and leg deflection, δ, as well as
the reaction force at the toe, F.
The raw benchtop data set has an uneven distri-
bution over the leg configuration space and contains
configurations that cannot occur during locomotion.
Therefore, prior to the fitting studies the raw data is
avaraged out and evenly sampled, then filtered to pick
those “physically relevant” points that correspond to
ground contact with no slippage characterized by two
conditions: 1) positive ground reaction force, Fn > 0;
and 2) surface force that is less than the maximum
Columb friction force, Fs < Fc . The maximum
Columb friction force is given by Fc = µFn where
Fn is the normal component of the ground reaction
force, and µ is the friction coefficient which is empri-
cally measured to be 0.5 for the 4-bar legs over card
board paper. Figure 4(left) shows the scatter-plot of
v ¯¯ ¯¯ 2
u ¯¯
u ¯¯d − dˆ¯¯¯¯
t
ˆ :=
p(d, d) 2
× 100
M (dmax )2
where M is the length of the data vectors and dmax
is the maximum value in original data.
Table 1 reports fitting and cross validation perfor-
mances for the leg model in (1) with varying polyno-
mial order, N . Since the improvement in model perfor-
mance is insignificant for quadratic and higher order
polynomials we choose to use the linear model due to
its simplicity. Figure 4 also presents the best linear
fits for the leg length, l, and deflection angle, δ.
4.2 Verification of Leg Model in Walk-
ing
In order to assess the performance of the linear leg
model during walking we ran a set of experiments
where the leg configuration as measured by a ”ground
truth” visual test station is compared with the leg
model output. During these experiments we concen-
trate on a single leg whose motion is recorded by a
high speed camera (RadLake HR-1000) at 125Hz and
processed off-line to obtain ground truth leg configu-
ration.

3
 Physical Relevant Data Set Affine Model for Leg Length Affine Model for Deflection Angle
30 160 8

7
25
155
6

20
150 5

15
δy

δ
l
145
3
10

140 2
5

1
135
0
0

δx σ σ
−5 130 −1
−10 −5 0 5 10 15 20 1.5 1.75 2 2.25 2.5 2.75 3 1.5 1.75 2 2.25 2.5 2.75 3

Figure 4: (Left) Scatter-plot of toe positions of the physically relevant cases. The axes of the plots are the cartesian
displacement of the toe point, (δx, δy), from the rest position which is indicated by the black circle. Resulting
scatter-plots for leg length, l, (center) and deflection angle, δ, (right). The best fit line is depicted on both plots.

Speed 0.2 m/sec 0.45 m/sec preliminary verification studies in Section 5.2.
p(l, ˆl) p(δ, δ̂) p(l, ˆl) p(δ, δ̂)
Exp #1 1.29% 10.34% 1.25% 15.00%
3
Exp #2 1.19% 7.77% 1.11% 16.95%
Exp #3 1.29% 8.77% 1.12% 12.01% B 2

Exp #4 1.14% 6.38% 1.11% 12.97% 1

H q3
Exp #5 1.29% 8.89% 1.23% 9.61% w3
Q
W s2 s0 q2
w1 w2 s1 q1
Table 2: Linear leg model performance during tri-
pod walking measured by the percentage root mean G
square error between the model output and visually 2 3
Ground
measured leg configuration. Small values indicate suc- 1 Truth
3 Measurement 1 2
cessful model prediction. System

Figure 5 shows typical comparisons of the visually

measured leg configuration states and the output of Figure 6: Robot supported on three legs on a flat
the leg model during a typical slow tripod walking gait ground surface, G. Two cameras track three mark-
(0.2 m/sec). The model performance for five runs at ers on the body which are depicted by numbered pen-
two different speeds can be found in Table 2 where the tagons. An off-line system computes body pose with
first two columns correspond to the plots in Figure 5. respect to the world coordinate frame, W.

5 Body Posture Computation 5.1 Computation of Body Pose within

a Tripod Stride
Define “tripod stride” to be a motion of the robot
where the body is supported by three legs whose Figure 6 illustrates a typical robot posture during a tri-
ground contact points (toes) are non-collinear and sta- pod stride. Without any loss of generality we will as-
tionary (no slippage and no lift-off) for a time inter- sume that supporting leg indices are i = 0, 1, 2. Using
val, t ∈ [0, ts ]. During a tripod stride the body frame, the leg model defined in (1) we compute the position
B, is related to the world frame, W, by a unique ho- of the ith toe, ci = m(σi ), in its associated hip clamp
mogenous transformation which is parameterized by frame, Ci , as a function of the strain across its compli-
the configurations of the supporting legs. ant part, σi . The hip clamp frame, Ci , is related to the
Driven by this observation we introduce an on-line body coordinate frame, B, by a homogenous transfor-
algorithm for 6-DOF body pose computation within a mation, hθi i : Ci → B, which is parametrized by the
tripod stride in Section 5.1 followed by a discussion of angular position of its hip shaft, θi . It directly follows

4
 Leg Length Comparison Deflection Angle Comparison
170
6
160
4

δ
l
150
2
140
0
130
1.2 1.3 1.4 1.5 1.6 1.7 1.2 1.3 1.4 1.5 1.6 1.7
170
6
160
4

δ
l

150
2
140
0
130
1.3 1.4 1.5 1.6 1.7 1.8 1.3 1.4 1.5 1.6 1.7 1.8
170
6
160
4

δ
l

150
2
140
0
130
1.2 1.3 1.4 1.5 1.6 1.7 1.2 1.3 1.4 1.5 1.6 1.7
170
6
160
4

δ
l

150
2
140
0
130
1.3 1.4 1.5 1.6 1.7 1.8 1.3 1.4 1.5 1.6 1.7 1.8
170
6
160
4

δ
l

150
2
140
0
130
1.2 1.3 1.4 1.5 1.6 1.7 1.2 1.3 1.4 1.5 1.6 1.7
t t

Figure 5: Comparison of visually measured (solid) and leg model generated (dashed) leg states during tripod
walking at 0.2 m/sec. Plots from five runs for the effective leg length, l, (top) and deflection angle, δ, (bottom)
are provided. Rows are ordered according to the experiment number to match with Table 2.

that the position of the ith toe in the body coordinate µ ¶

frame, si ∈ B, is given by si (θi , σi ) = hθi i ◦ m(σi ). α := arctan √ π3 ◦b1
Using the three ground contact points, si , we define µ (π1 ◦b1 )2 +(π2 ◦b1 )2
¶
the support coordinate frame, Q, centered at s0 with γ := arctan √ π3 ◦b2
(5)
(π1 ◦b2 )2 +(π2 ◦b2 )2
an orthonormal basis, µ ¶
√ π1 ◦b1
β := arctan ,
(π2 ◦b1 )2 +(π3 ◦b1 )2
q1 := e
£ 1, ¤ ¯¯ ¯¯
q2 := e2 − (qT1 e2 )q1 / ¯¯e2 − (qT1 e2 )q1 ¯¯2 , where we use the notation πi ◦f to denote the ith entry
q3 := q1 × q2 , of a vector, f ∈ Rn .
(2) Note that during a stride the ground contacts re-
represented in the body frame, B, where e1 := (s1 − main stationary, hence, the support frame, Q, is rigidly
s0 )/ ||s1 − s0 ||2 and e2 := (s2 − s0 )/ ||s2 − s0 ||2 are attached to the world coordinate frame, W, and re-
the unit vectors along the two edges of the support lated by a homogenous transformation, r : Q → W,
triangle. A homogenous coordinate transformation, which can be specified by the inital pose of the the
r : B → Q, relates the body frame, B, to this support body at the begining of the stride.
frame, Q,
5.2 Emprical Results
r(b) := Ab − As0 , (3)
T To assess the performance of the body pose compu-
where A := [q1 q2 q3 ] .
tation algorithm on an operational robot we compare
The 6-DOF configuration of the body in the support
the ground truth measurements of the body pose to
frame, Q, is composed of the position of the center of
the output of this algorithm.
mass (COM),
We attach three markers on the body denoting the
body frame, B, as illustrated in Figure 6. During lo-
u := −As0 , (4)
comotion a visual tracking system [7] logs the marker
and three rotational configurations: roll, γ; pitch, α; trajectories which are employed by an off-line process
and yaw, β. The body lateral and fore/aft unity vec- yielding 6-DOF body posture in world frame, W.
T
tors in the world coordinates are b1 = A [1 0 0] and Because the pose model represented in equations (4)
T
b2 = A [0 1 0] , respectively. By direct computation and (5) is predicated upon the assumption that the
we obtain the rotational configurations, robot is supported by a single tripod, it is important to

5
 2
prical verification of the model both in benchtop ex-
periments and in the normally functioning robot are
x
0

−2 presented where we demonstrated less than 2% error in

2 2.2 2.4 2.6 2.8 3 3.2 3.4 3.6 3.8 4
10 leg length and less than 10% error in angular position
y

5
of toe.
0
5
2 2.2 2.4 2.6 2.8 3 3.2 3.4 3.6 3.8 4 Based on the leg configuration model an algorithm
for computation of robot posture during tripod sup-
z

−5
port is introduced. Preliminary experimental data
2 2.2 2.4 2.6 2.8 3 3.2 3.4 3.6 3.8 4
5 comparing its performance against ground truth mea-
α

0 surements is presented.
−5
2 2.2 2.4 2.6 2.8 3 3.2 3.4 3.6 3.8 4
In the near future we plan to fuse other sensory
5
data streams such as 3-DOF laser rate gyros readings
γ

0
and visually registrated landmarks to build a full body
−5
2
2 2.2 2.4 2.6 2.8 3 3.2 3.4 3.6 3.8 4
state estimator mechanism with well characterized er-
ror margins. Such a technology will enable us to imple-
β

−2
2 2.2 2.4 2.6 2.8 3 3.2 3.4 3.6 3.8 4
ment dynamical feedback controllers with significantly
t improved robustness and efficiency.

Figure 7: Pose states of the rigid body (solid) mea- 7 Acknowledgements

sured by GTMS and (dashed) computed according to
our algoritm.
extract cues from the strain measurements that mark
the onset and termination of the tripod stride. To this
end, taking advantage of the assumption of a level sup-
port surface, we define a measure of co-planarity for
P
6 P
6 References
all six legs, ρ := sTi si − s T s where s := 1/6 si .
i=1
Smaller values of ρ imply that all six toes are close to
a common plane. In tripod walking small ρ values in-
dicate double stance, and large values point out tripod
support. We emprically pick a threshold, ρ̂, to deter-
mine the start and the end of tripod support intervals.
Figure 7 compares the body pose states obtained from
ground truth measurement and output of pose com-
putation algorithm showing very close match in most
variables.
The reader should notice that the visual tracking
system presently has a 5mm mean accuracy — coarse
enough to jeopardize the accuracy of the angular state
measurements. Furthermore, since both cameras of
the tracking system look at the sagittal plane those
We thank Richard Groff for his careful contributions
to the modeling effort and Gregory Sharp for his vi-
sual tracking system implementation. This work is
supported by DARPA/SPAWAR Contract N66001-00-
C-8026.
[1] H Komsuoglu. Legnet - a distributed wireless network. Technical
i=1 report, University of Michigan, 2002.
[2] H. Komsuoglu, D. McMordie, U. Saranli, N. Moore, M. Buehler,
and D. E. Koditschek. Proprioception based behavioral ad-
vances in a hexapod robot. In International Conference on
Robotics and Automation, volume 4, pages 3650–3655, Seoul,
Korea, 2001.
[3] P-C. Lin. Benchtop setting for compliant leg test and modeling.
Technical report, University of Michigan, 2003. In preperation.
[4] E.Z Moore. Leg design and stair climbing control for the rhex
robotic hexapod. Master’s thesis, McGill University, November
2001.
[5] U. Saranli. Dynamic Locomotion in a Hexapod Robot. PhD
thesis, Univerisity of Michigan, 2002.
[6] U. Saranli and D. E. Koditschek. Back flips with a hexapo-
dal robot. In IEEE International Conference on Robotics and
Automation, volume 3, pages 2209–2215, 2002.

states with significant component orthogonal to the [7] G. Sharp. Ground truth measurement system, 2003.
sagittal plane, such as x displacement, have relatively http://sourceforge.net/projects/gtms/.

higher error. [8] Saranli U., M. Buehler, and D. E. Koditschek. Rhex - a simple
and highly mobile hexapod robot. International Journal of
Robotics Research, 20:616–631, 2001.

[9] Joel D. Weingarten, Martin Buehler, Richard E. Groff, and

6 Conclusion Daniel E. Koditschek. Automated gait generation and optimiza-
tion for legged robots. 2003. In preperation.
We have introduced a quasi-static leg model for a 4-
bar linkage leg structure relating the strain in its rear
compliant member to its configuration. Detailed em-